• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2019 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14
15import numpy as np
16
17import mindspore as ms
18import mindspore.nn as nn
19from mindspore import Tensor, context
20from mindspore.common.parameter import Parameter
21from mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
22from mindspore.nn.optim.momentum import Momentum
23from mindspore.ops import operations as P
24from mindspore.train import Model
25from mindspore.context import ParallelMode
26from tests.dataset_mock import MindData
27
28
29class Dataset(MindData):
30    def __init__(self, predict, label, length=3):
31        super(Dataset, self).__init__(size=length)
32        self.predict = predict
33        self.label = label
34        self.index = 0
35        self.length = length
36
37    def __iter__(self):
38        return self
39
40    def __next__(self):
41        if self.index >= self.length:
42            raise StopIteration
43        self.index += 1
44        return self.predict, self.label
45
46    def reset(self):
47        self.index = 0
48
49
50class TransposeNet(nn.Cell):
51    def __init__(self, strategy1, strategy2):
52        super(TransposeNet, self).__init__()
53        self.matmul = P.MatMul().shard(((8, 1), (1, 1)))
54        self.matmul_weight = Parameter(Tensor(np.ones([128, 256]), dtype=ms.float32), name="weight")
55        self.transpose1 = P.Transpose().shard(strategy1)
56        self.transpose2 = P.Transpose().shard(strategy2)
57
58    def construct(self, x):
59        x = self.matmul(x, self.matmul_weight)
60        x = self.transpose1(x, (1, 0))
61        x = self.transpose2(x, (1, 0))
62        return x
63
64
65def transpose_net(strategy1, strategy2):
66    return TransposeNet(strategy1=strategy1, strategy2=strategy2)
67
68
69def transpose_common(strategy1, strategy2):
70    learning_rate = 0.1
71    momentum = 0.9
72    epoch_size = 2
73
74    context.reset_auto_parallel_context()
75    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=8,
76                                      parameter_broadcast=False)
77
78    predict = Tensor(np.ones([32, 128]), dtype=ms.float32)
79    label = Tensor(np.ones([32]), dtype=ms.int32)
80    dataset = Dataset(predict, label, 2)
81    net = transpose_net(strategy1, strategy2)
82
83    loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
84    loss.softmax_cross_entropy.shard(((8, 1), (8, 1)))
85    opt = Momentum(net.trainable_params(), learning_rate, momentum)
86    context.set_context(mode=context.GRAPH_MODE)
87    model = Model(net, loss, opt)
88
89    model.train(epoch_size, dataset, dataset_sink_mode=False)
90
91
92def test_transpose1():
93    strategy1 = ((1, 8),)
94    strategy2 = ((1, 8),)
95    transpose_common(strategy1, strategy2)
96
97
98def test_transpose2():
99    strategy1 = ((1, 4),)
100    strategy2 = ((1, 8),)
101    transpose_common(strategy1, strategy2)
102
103
104if __name__ == '__main__':
105    test_transpose1()
106    test_transpose2()
107